Hi everyone!
I'm working on a multitemporal classification of the forest types from the Sentinel data. The reference data for each class is represented by multiple polygons scattered over the study area. After I overlay polygons and satellite data (88 S2 tiles from different period of year) my training dataset becomes really huge and I'm looking for a way to reduce its size.
I'm thinking of doing aggregation by polygon and date, and so far I have calculated mean value, max, min and some other quantiles. The cross-validation suggest that I'm not that far for when I'm using the complete dataset.
Does anyone know some better way or can provide me useful references how to perform aggregation per polygon?
Or should I completely abandon this approach and use some other method for reduction of data size?
Thanks in advance!
Extraction of significant statistical values can be used to reduce the volume of the data. Maybe one point worth remembering in case you are aiming at deploying the classifier as part of a production process, is that it means that after training your classifier these aggregation/reduction methods will have to be applied on any future input data, which can also have consequences in terms of processing time/performance. To summarize, always think what your real input data will look like and create a classifier suitable for this type of data.
Dear Jovan Kovačević
It depends upon the type of data you choose or create.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training